making tougher ai opponents

minipol · #1 September 10th, 2002, 08:42 AM

Hi,

recently a number of threads were dedicated to how the ai behaves and what he uses. If you would take out all of the techs he doesn't use, wouldn't that improve the ai? For instance, it doens't use robo miners so if you switch that of in the research tree, it will not research it (which can be prevented with modding) but it also prevents the user from getting an advantage over the ai because it simply isn't there. If you would do that to all other technologies the ai never uses, wouldn't that make the ai a bit tougher?

[ September 10, 2002, 07:42: Message edited by: minipol ]

dogscoff · #2 September 10th, 2002, 09:59 AM

It won't make the AI tougher but it will take away some of the tools to beat it.

The thing is though, the real problem is in the AI's predictable behaviour, complete lack of any understanding of what's going on around it and total inability to make plans and follow them through.

These things aren't in the tech tree, and if they were I wouldn't want to mod them out of myself=-)

Don't get me wrong BTW. I often play vsAI and I really enjoy it. The AI is no match for a human intellect (altohugh with bonuses and a good start it can out-produce one and provide a challenge.), but it's still excellent compared to some other game AIs I've seen.

minipol · #3 September 10th, 2002, 12:46 PM

indeed it would take away some tools to beat it. i wonder how one could make an adequate ai for such a complex game. i often wondered how people make an ai for such a game. I would go for the chess engine approach of generating possible actions and assigning a weight to them. would be very tough but a lot of those algorithms are already described. still, not a trivial task

Makinus · #4 September 11th, 2002, 01:42 AM

One way to improve the AI in the game would be to change the way the AI uses the planets it colonizes...

The AI does a very poor use of the planets it colonizes, like building Mineral Extracors in mineral poor planets while Research Facilities would be more useful... (this often hapeen when the AI set a colony to Contruction Yard).

I tried to mod additional colony types for the AI, but looks like the types the AI recognizes are hard coded in the EXE and can't be changed only modding the TXT files...

You can only work with the colony types provided (8 types if i'm not wrong), and this is not enough to create an effective strategy for colonization...

I hope that MM/Shrapnel changes this in the next patch, allowing additional colony types (just changing it from hard coded to the text files will do), so i can create a more chalenging AI....

Well, I contributed my two cents to the topic.... it's time of the rest of you guys to contribute your ideas for a more chalenging AI....

Makinus

dogscoff · #5 September 11th, 2002, 01:59 AM

I think the best way to improve SE4's AI is to meaningfully increase the complexity of the AI's decision-making processes. The more complex a system is and the more variables it responds to, the harder it is to predict its response to the stimuli you give it.

Make those responses meaningful rather than random, and you'll have a pretty intelligent looking AI. See the thread a while back which discussed "nethack", a game which has undergone this process for the Last 20 years or so, and is now (apparently) frighteningly clever.

As an analogy, consider the combat modifiers. At the base of it, you have a simple calculation: 100% chance of hitting, and then deduct 10% for each square of range.
Too simple. Too predictable. So the system was made more complex: modifiers were added for hull size, racial bonuses, your sensors & enemy ECM, your experience, their experience, war shrines... Now, in order to predict the outcome of a ship firing on another, you have to know all those different variables. We have acheived unpredictability- except that the combat modifier system is pretty much transparent, with all the to-hit percentages readily available at the click of an icon.

Imagine if it wasn't though. Imagine that you couldn't see your chance to hit, or the modifier you get from experience or sensors or ECM or anything. You would have an unpredictable system, but the important thing is that it would be meaningful unpredicatbility (good) rather than completely random unpredictability (bad). Why meaningful? Because the unpredictable elements are governed by meaningful decisions. (ie sensors increase hit rates, ECM decrease them) The trick to unpredictability is making these decisions so numerous and based on so many different things that you can't hope to accurately predict their combined result.

And that is the way forward with the AI. For example:
in early Versions of SE4, the AI would completely ignore minefelds and throw ship after ship into them. Then MM improved the system slightly and made it so that when the AI first encounters mines, it immediately researches minesweepers. However, this is a simple and predictable equation that can be exploited by a human - drop one single mine and watch the AI waste research and resources developping minesweepers. What is needed now is to make the system more complex, so that it becomes less and less predictable: the AI could look at the importance of the mine-laying empire's territory when deciding whether or not to invest in minesweepers. If it's some backwater neutral, don't bother. Does it already have the pre-requisite techs? Maybe also have the decision affected by the urgency of other research commitments and the possibility of other solutions (making peace, developping armour, expanding in a different direction.) By the time all this was added, you'd have absolutely no idea how the AI will respond if you drop a mine in its path. You will know, however, that it will make a pretty sensible decision.

Balancing these decisions won't be easy, but that's what modders are for. As these levels of complexity are added in, bit by bit, patch by patch, the AI will get exponentially smarter and smarter.

This seems to be more or less what MM is gradually doing now with SE4, although I get the feeling (and this is just my hunch, I have no real evidence) that SEIV patching is rolling towards a halt, concentrating more now on fixing bugs for a final, bug-free Version of SEV and Version 1.00 of SEV.

The only trouble with this method of development is that it's a very lengthy process, and that we have only one programmer. We'll get there eventually (Space Empire XII?), but patience will be required...